Particle Image Velocimetry (PIV) has been used successfully for measuring instantaneous two dimensional velocity fields. Analyzing PIV images involves matching particle images captured sequentially. In the usual practice, correlation (auto- or cross-correlation) is used to find the displacement (hence velocity) of the particles within a large number of small 'interrogation areas' in the field of view. An image correlation within an interrogation area is inherently sensitive to the rotation and distortion of the fluid flow within the area, and is also sensitive to the brightness of particles. It can only find the mean spatial offset of particles weighted by their brightness. Rotation, dilation and distortion of fluid flow within the interrogation region and intensity changes of particles introduced correlation errors or bias. As a consequence, the size of the interrogation area and the time interval between images must be kept small to avoid these problems. A feature-recognition method is proposed here for analyzing PIV images. It first extracts structural features of the particle pattern after their locations have been isolated from images. A preliminary process is to replace the particle images by the Cartesian coordinates of particle centers. In this way the brightness of particle images plays no further part, and the point positions are used to establish structural features: topological relations between each point and its neighbors. The interrogation area is defined by a limited number of neighboring points. The size and shape of each interrogation area varies with the distribution of neighbors. A fit to motion, rotation and distortion among the neighbors is then carried out in the space of topological relations of successive images. In this way changes of structural features define fluid spatial translation, rotation, and deformations within each interrogation region. Measurement of feature space in two successive images demands knowledge of the locations of corresponding points derived from individual particles in the two images. Classification of point correspondences, despite confusingly discordant displacements from one image to the next, can be made by taking advantage of physical limitations on the possible movement of particles between the two images. It is found that feature space search and correlation is a much more efficient procedure than correlation operations in the two dimensional image domain.